Research Seminar
A numerical convex hull based procedure for selecting among multilevel component solutionsEva CeulemansKU Leuven | |
| Abstract: | Given two-level multivariate data - for instance, time series data of different subjects - one often wants to gain an understanding of the underlying latent processes. To this end, Timmerman (2006) proposed a class of multilevel component models. These models consist of a separate component model for each level in the data. Specifically, the differences between the subjects are captured by a between-component model and the differences within the data of each subject by a within-component model. Within the class of multilevel component models a number of variants can be distinguished. These variants differ with respect to the within-component model only, in that different sets of restrictions are imposed on the within-component loadings of the variables and on the variances and correlations of the within-component scores of the measurement occasions. When applying multilevel component analysis in practice, it is usually unclear beforehand which model variant with which number of between and within components should be fitted to the data. In most cases, this uncertainty is dealt with by fitting the different model variants with varying numbers of between and within components and by applying a model selection procedure. The procedure that is almost always used does not take differences in restrictiveness of the different multilevel component analysis variants properly into account, however. In this paper, we will present a formal model selection procedure in which these differences in restrictiveness are reflected in the complexity measures of the obtained solutions. Specifically, building on the work of Ceulemans and Kiers (2006, 2009) for the related family of three-mode component analysis models, we propose a numerical convex hull based model selection procedure, which selects the solution with the best balance between number of free parameters and goodness-of-fit. The performance of this hull method is systematically evaluated in a simulation study. |
| Date: | Tue Nov 17, 12:15 pm - 1:15 pm |
| Place: | room 02.51 (Department of Psychology, Tiensestraat 102, 3000 Leuven) |